Detect and Defense Against Adversarial Examples in Deep Learning - presented by Assoc. Prof. Wassim Hamidouche

Detect and Defense Against Adversarial Examples in Deep Learning

Assoc. Prof. Wassim Hamidouche

Assoc. Prof. Wassim Hamidouche
Slide at 21:17
Adaptive denoising with the block matching 3D (BM3D) filter [10] 1. Image patches clustering P(P) = (10) [14] Dabov, K. et al., Image denoising by sparse 3-d transform-domain collaborative filtering, IEEE Transactions on image processing16(8),2080-2095 (2007) W. Hamidouche - Proposed approach 0000000 20/32
1
References
  • 1.
    K. Dabov et al. (2007) Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering. IEEE Transactions on Image Processing
Share slide
Summary (AI generated)

So, what we do here is use a block matching 3D filter. The first step involves clustering different patches of the image. We compute the distance between two patches, Q and P, and if the distance is below a certain threshold, we group all these patches into one cluster, denoted as π. This forms the first step, where we create 3D piles of patches that are similar in terms of distance.

Next, we apply a 3D linear transform, which is provided here. This transform helps us to convert the piles into the transform domain. Within this domain, we perform a simple shrinkage technique to remove all coefficients that are lower than a specific threshold. This process is referred to as threshold link shrinkage.

Overall, our approach involves using a block matching 3D filter to cluster patches, compute distances